Skip to main content
Open the Results tab at any time. You never have to wait for collection to finish.

The verdict

Results lead with a verdict card: the current leader (“Warm-v2 wins”), the confidence score, the vote count, prompt-group coverage, and a plain next step such as “Collect 12 more votes before finalizing” or “Signal is strong enough to finalize.” When you are satisfied, select Complete evaluation to lock in the answer. Completed evaluations stay readable and exportable.

What confidence means

HeyBee continually tests its current inference against the newest judgments. Pairwise criteria use preference outcomes. Scored criteria also use score magnitude and paired score differences, while correcting for voter scale use after enough observations. Confidence rises only when held-out evidence remains consistent. For scored evaluations, Results also shows normalized scores, uncertainty, probability of being best, the meaningful margin, score distributions, and the state of every critical criterion. A relative winner can still be reported when all candidates score poorly, but HeyBee displays a low-absolute-quality warning. A result asks for more collection until it has at least 10 counted votes covering at least half of your prompt groups. “Signal is strong enough to finalize” appears at a score of 80 or higher.

Rankings

The ranking table lists every candidate with its win rate and vote count. With multiple criteria, each criterion has its own tab and leaderboard, so you can see that one candidate wins on Detail while another wins on Motion. Settings evaluations add a Decision confidence figure, the estimated probability that the recommended configuration beats a typical one, and a What matters breakdown that calls each setting Solid, Likely, or Tentative based on how clearly the votes favor one value.

Behind the result

The Reliability tab holds the evidence: every individual vote with any written voter notes, consistency signals, and per-voter reliability. Follow See what’s behind this on the verdict card to land there directly. Results are visible only to the evaluation owner and Shared workspace members whose role allows it. HeyBee never publishes results publicly. Exports contain every raw preference and everything needed to reproduce the decision. That makes them dual-purpose: the same file that backs your decision is a preference dataset, ready for RLHF or DPO training. Exports include voter feedback, so treat them as sensitive data.